It stands for Importance Matrix, a technique used to improve the quality of quantized models.
The Imatrix is calculated based on calibration data, and it helps determine the importance of different model activations during the quantization process.
The idea is to preserve the most important information during quantization, which can help reduce the loss of model performance, especially when the calibration data is diverse.
[1][2]
For imatrix data generation, kalomaze's groups_merged.txt with added roleplay chats was used, you can find it here. This was just to add a bit more diversity to the data.
Officially rebranding Thespis to TheSpice. Why? Cause its a cooler, simpler name.
I've focused on making the model more flexible and provide a more unique experience.
I'm still working on cleaning up my dataset, but I've shrunken it down a lot to focus on a "less is more" approach.
This is ultimate a return to form of the way I used to train Thespis, with more of a focus on a small hand edited dataset.
Datasets Used
Dolphin
Ultrachat
Capybara
Augmental
ToxicQA
Yahoo Answers
Airoboros 3.1
Features
Narration
If you request information on objects or characters in the scene, the model will narrate it to you. Most of the time, without moving the story forward.
You can look at anything mostly as long as you end it with "What do I see?"
image/png
You can also request to know what a character is thinking or planning.
image/png
You can ask for a quick summary on the character as well.
image/png
Before continuing the conversation as normal.
image/png
Prompt Format: Chat ( The default Ooba template and Silly Tavern Template )
image/png
If you're using Ooba in verbose mode as a server, you can check if you're console is logging something that looks like this.
